Learning by Fixing: Solving Math Word Problems with Weak Supervision

• DOI: 10.1609/aaai.v35i6.16629
• Publication Date: 2022-09-08T18:44:31Z
• Authors: Yining Hong; Qing Li; Daniel Ciao; Siyuan Huang; Song-Chun Zhu
• Keywords: Regularization; Tree (set theory); Root (linguistics); Supervised Learning

Previous neural solvers of math word problems (MWPs) are learned with full supervision and fail to generate diverse solutions. In this paper, we address issue by introducing a weakly-supervised paradigm for learning MWPs. Our method only requires the annotations final answers can various solutions single problem. To boost learning, propose novel learning-by-fixing (LBF) framework, which corrects misperceptions network via symbolic reasoning. Specifically, an incorrect solution tree generated network, fixing mechanism propagates error from root node leaf nodes infers most probable fix that be executed get desired answer. more solutions, regularization is applied guide efficient shrinkage exploration space, memory buffer designed track save discovered fixes each Experimental results on Math23K dataset show proposed LBF framework significantly outperforms reinforcement baselines in learning. Furthermore, it achieves comparable top-1 much better top-3/5 answer accuracies than fully-supervised methods, demonstrating its strength producing